Inside LimX Dynamics: The Honest Truth About Humanoids

[00:00:01][music] >> Uh we are at Limex. Is that how you say it? Limex?

[00:00:07]>> Limex.

[00:00:07]>> Okay, Limex Dynamics in Shenzhen.

[00:00:10]And uh we are joined by >> [music] >> Genie.

[00:00:12]>> Genie. So, you're going to show us some robots?

[00:00:15]>> Yes, uh >> [music] >> very glad to have you all here and welcome to Shenzhen and Limex. Um so, we are going to show you some of our robots, humanoid robots, and our bipedal dual robots.

[00:00:34]We all focus on [music] general-purpose robots.

[00:00:38]>> Okay.

[00:00:38]>> So, we develop robots in two very different way.

[00:00:42]The first way is we build full-size humanoid.

[00:00:46]>> Mhm.

[00:00:46]>> So, we don't change the form factor of the [music] robot.

[00:00:50]>> Right.

[00:00:50]>> But, it can be adapted to most of the scenarios or applications.

[00:00:55]>> Mhm.

[00:00:55]>> And on the other hand, we we are focused [music] on module design robots. So, you can change the form factor of the robots to adapt to >> [music] >> different vertical applications.

[00:01:08]>> Right.

[00:01:08]>> So, very different ways, but the ultimate goal is for general purpose.

[00:01:14][music] >> How are you, Ollie?

[00:01:19]He He replied He replied in Hindi.

[00:01:33]How's the weather today, Ollie?

[00:01:37]>> So, you can Hey, hold Hold on, Ollie.

[00:01:38]Hold on.

[00:01:40]Hold on, Ollie.

[00:01:41]>> Just go on, Ollie.

[00:01:42]>> So, we we are Yeah, we we because sometimes it talks too much.

[00:01:50]So, we we also have Ollie to stop.

[00:01:54]And also, We want Ollie's response to be very quick and make sure feel very comfortable. [music] It's just like talking to a human.

[00:02:03]>> Yeah.

[00:02:03]>> So, let's take a look at our tool.

[00:02:05]>> Tron.

[00:02:05]>> They.

[00:02:14]Okay.

[00:02:15]Yeah.

[00:02:16]>> Very cool.

[00:02:17]>> Look at me. And And this is Tron one, the first generation. So, it's a bypedal robot.

[00:02:22]>> Mhm.

[00:02:23]>> You buy one product and then you can have three different kinds of foot end.

[00:02:27]The wheel base, the sole base, [music] and the point foot.

[00:02:31]>> I see.

[00:02:32]>> Yeah. Um you can [music] like try to put your hand on top and push as hard as you can.

[00:02:38]Try it.

[00:02:40]>> Wow. It's like a suspension dynamic [music] suspension.

[00:02:45]>> Yes.

[00:02:46]>> Nice.

[00:02:49]>> So, this is um and we can have uh Tron [music] product. Yeah.

[00:02:54]>> Let Let me.

[00:02:57]Don't sit on it.

[00:03:00]>> Oh, it's quite strong.

[00:03:01]>> Yes.

[00:03:03]>> Very stable.

[00:03:04]>> So, it has uh totally eight joints.

[00:03:06][music] The payload is about 10 kg. Not that much, but you still can feel the power [music] of the the the robot.

[00:03:19]>> [music] [music] >> We are joined by Paul Lee. He is uh heading R&D at LimX Dynamics. Thank you so much for joining the Humanoid Hub.

[00:03:42]>> yeah. My pleasure.

[00:03:43]>> We saw the robots doing amazing things here. To start with, I wanted [music] to know about your background and how you started uh how you how you started getting interested in humanoids?

[00:03:53]>> All right. All right.

[00:03:55]I'm an engineer by training. I got my PhD from the states University of Illinois [music] and Urbana-Champaign. I'm in mechanical engineering.

[00:04:04]>> Okay.

[00:04:04]>> So, like I have like basic understanding of all these mechanical things and control stuff. [music] And afterward I work on various sectors of industry, aerospace, you know. Like consumer electronics, things like that.

[00:04:21]Four five years ago, I started to work on this the precursor of this company with the founder.

[00:04:30]>> Okay.

[00:04:30]>> That's the opportunity I got to to work on a humanoid. At that time it was not humanoid is it was quadruped.

[00:04:38]>> I see.

[00:04:39]>> And then later on we realized with all the evolution of the methods it was doable.

[00:04:46]>> Right.

[00:04:47]>> For the hardware and the software and the algorithms the humanoid it was doable. And then like three years ago we started like all in on the humanoid [music] robot. And then right now we have a several robots on the ground already.

[00:05:01]>> Okay.

[00:05:02]So, what's your core philosophy behind pursuing the form factor humanoid form factor? Why is it so important?

[00:05:08]>> It's important because it has the you know the feature of working with the human beings.

[00:05:15]>> Mhm.

[00:05:15]>> Yeah, so humanoid and human beings working together that's the philosophy.

[00:05:19]>> Okay.

[00:05:19]>> Right. So, but it has to be able to uh situated in the human environment.

[00:05:26]>> Mhm.

[00:05:26]>> So, that's our basic understanding of how humanoid could be useful. Even though it could be sub optimal.

[00:05:34]>> Right.

[00:05:34]>> That's why we also have another product series called Tron which is not like 100% humanoid.

[00:05:41]>> Right.

[00:05:41]>> We believe it's it's a solution. It might be a optimal solution [music] better than humanoid.

[00:05:47]>> Mhm.

[00:05:48]>> So like we are pushing two front >> Right.

[00:05:50]>> at the same time.

[00:05:51]>> Okay.

[00:05:52]And we saw a lot of dynamic locomotion with Ali and Tron.

[00:05:57]>> Yes.

[00:05:57]>> Like it's the balance is amazing and it can do like climb stairs and stuff.

[00:06:03]So what were the core challenges and solving locomotion and get get to such a good level?

[00:06:09]>> Oh yes.

[00:06:10]So I think three years ago like I just mentioned the time we realized humanoid locomotion could be solved with the advancement of reinforcement learning.

[00:06:21]>> Okay.

[00:06:22]>> Together with like modeling of the dynamics in a parallel fashion.

[00:06:27]>> [music] >> I see.

[00:06:28]>> it enabled large scale training.

[00:06:31]>> Right.

[00:06:32]>> So the secret ingredient now it's not secret anymore is the reinforcement learning and the both for Tron and the humanoid.

[00:06:40]>> Mhm.

[00:06:41]>> But the challenges are there.

[00:06:42]>> Right.

[00:06:44]>> Especially if you want to put the the product in [music] the product level.

[00:06:48]Reliability and stability like that.

[00:06:51]>> I see.

[00:06:51]>> [music] >> Three years ago we started realize that it was robust because through training it was a like highly non-linear controller that was able to >> [music] >> cover different working scenarios which was like huge advancement from the previous model predictive controls which was supposed to have only one model and you [music] know surrounding areas as a working uh areas for the humanoid to perform.

[00:07:21][music] With reinforcement learning we we see that the non-linearity can encompass [music] the different working modes. And then like I just mentioned the reliability was the issue because we really want to [music] uh close the sim to real gap.

[00:07:37]>> Okay.

[00:07:37]>> That was the key to achieve the high reliability because um >> [music] >> in a simulation you can always see it so like working perfectly like working all the time but >> [music] >> when when it comes to real deployment >> Mhm.

[00:07:52]>> different terrains you really see the challenges from simulation to the reality.

[00:07:58]>> Okay.

[00:07:58]>> both on the body of the humanoid ultron and the interaction between the robots and the environment. So we have to simulation capture both.

[00:08:10]>> Okay.

[00:08:11]>> So so it achieves the performance intended.

[00:08:15]>> I see.

[00:08:15]>> So that's the thing. The other challenge is from the hardware.

[00:08:19]Okay. So our founding team was basically from algorithm world, you know.

[00:08:26]>> We we did algorithm for like more than 10 years [music] both in academia and industry. Myself industry founders >> [music] >> well John from academia.

[00:08:36]But we realized that at that time we didn't have a good platform to really like perform the algorithm research and development. So that's why we need to develop the hardware from scratch [music] even though we're not that you know familiar with the hardware building but we realize that the only way to go and the challenge will be provide hardware >> [music] >> that was good enough to be able to put the algorithm on top of them.

[00:09:05]So that's another challenge because the hardware involves a lot of small parts supply chains and you have different [music] industry per se because you have to connect to the you know automobile industry to give you the best gears.

[00:09:21]You have to uh you know consult to the consumer electronics to give the cost down a little.

[00:09:28]>> Right.

[00:09:28]>> So it's [music] it's all different field that's why it was a challenge posed for us and they were still in the process of overcoming the challenge.

[00:09:38]>> So it's it's more of a systems problem >> Yeah, yes.

[00:09:42]>> reinforcement learning hardware and uh closing the sim to real gap.

[00:09:46]>> Right, right.

[00:09:47]>> Yeah. All working together.

[00:09:48]>> All working together. It's a system integration.

[00:09:50]>> Yeah, yeah.

[00:09:51]>> Yeah.

[00:09:52]>> So in humanoids, especially the ones that are available in the market, I hear a lot of complaints about motors getting uh you know >> wear and tear.

[00:10:02]>> heated up. Right.

[00:10:03]>> Yeah, wear and tear.

[00:10:04]>> There are reliability issues, repair issues.

[00:10:07]>> Yes.

[00:10:09]>> So how are you going to differentiate in the performance and you know battery efficiency >> Mhm.

[00:10:14]>> uh and what are the core innovations behind [music] the mechanical you know and the the power >> Mhm.

[00:10:20]>> side?

[00:10:20]>> Well, uh I think one [clears throat] thing is like we we really got into the nitty-gritties of hardware building. We don't just procure you know the joints the motors from the other suppliers. We build them all ourselves.

[00:10:35]Uh that's why we understanding [music] we understand that all the details of this hardware so that we can put the details into the [music] simulation so that the like I said the sim to real gap >> Mhm.

[00:10:47]>> and then can be closed from that >> Right.

[00:10:50]>> Mhm.

[00:10:50]>> Uh so that's one thing. The other thing is >> [music] >> like you mentioned it's a system engineering.

[00:10:56]Uh not only you know the the actuators but how the actuators are connected to the computing boards.

[00:11:04]>> Mhm.

[00:11:04]>> How the computing boards are optimized so that uh >> [music] >> you know the thermal management can be done in the you know constrained space.

[00:11:12]>> Mhm.

[00:11:13]>> So all we have all the control of those variables so that in the end we have optimal solution and we're getting there. It's not there yet.

[00:11:21]>> Okay.

[00:11:22]>> So we're confident that with the you know [music] gradual honing the the process and you know like I mentioned from [music] down to the motor level to the algorithm level.

[00:11:32]>> I see.

[00:11:32]>> We will be able to achieve better reliability and stability and then ultimately getting this into the factory floor.

[00:11:41]>> Okay.

[00:11:42]And I wanted to ask about manipulation.

[00:11:44]We saw the teleop demo on prawn.

[00:11:48]So in your tech stack, how important is AI and what are you building in training pipeline or models? Are you doing What are you doing in that research side?

[00:11:58]>> All right. All right. Yeah. So our view of uh you know all these VOA models, you know, world [music] action models are that they have to be useful in the real environment.

[00:12:09]>> Mhm.

[00:12:10]>> We [music] as a company we don't like uh train the models for the sake of it.

[00:12:16]We really want to close loop with you know the users who want to really find values from those algorithms together with form factors.

[00:12:26]>> [music] >> So our approach is to work uh with you know the users.

[00:12:32]>> Mhm.

[00:12:32]>> They can be from the factories. They they they are they can be from [music] the you know R&D field but we work with them.

[00:12:39]>> Okay.

[00:12:40]>> we understand each each individual verticals.

[00:12:44]>> Right.

[00:12:45]>> Uh the real requirement.

[00:12:46]>> Mhm.

[00:12:47]>> We don't really want to just do academic work to push the [music] benchmarks per se. We really want to have the robot working on the factories [music] even though at the first it can be remote controlled.

[00:13:00]>> Right.

[00:13:00]>> So like I said remote control could be a very important stack in this front because [music] the efficiency because remote control at the first it will be the data collection tool.

[00:13:10]>> Right.

[00:13:10]>> And you really want the data collection tool to be very efficient. You don't want the you know 8 hours work [music] and only 2 hours of useful data. You really want to push useful data to be [music] like 6 hours.

[00:13:21]>> I see. Yeah. Yeah.

[00:13:22]>> So that that's our understanding that we really want to have a remote good teleop control system >> [music] >> for Tron and for humanoid as well.

[00:13:32]>> I see.

[00:13:33]>> So for Tron, it's a little bit easier because you only have [music] two arms. For humanoid only, you really need to have a whole body [music] control system, which is still a you know, frontier of the research in this field.

[00:13:47]>> [music] >> So we really we built our behavioral model for the Ollie so that it can perform a whole body movement to do the teleop teleop for we call it uh >> [music] >> uh you know, manipulation locomotion together. Locomanipulation.

[00:14:05]>> You have an amazing robot.

[00:14:08]All right.

[00:14:09]What do you think are the key limiting factors that will allow you to scale Ollie and be useful?

[00:14:18]>> Uh limiting factors are the things I just mentioned. I we're still optimizing the whole process. Each individual component to to make them work coherently [music] to achieve certain reliability. Right now it's still still in the process.

[00:14:34]It's clunky still because we don't haven't found optimal solution for you know, how the motors are >> [music] >> actuating according to the motions calculated from uh reinforcement learning algorithms. Very precise. So and also thermal management is still an issue. You see a lot of these holes for the you know, dissipations but still there might be better solutions because [music] with all these holes uh with with all these exposure of the internal just for the better dissipation, you lose a lot of the opportunities to put [music] this robot into certain environment requires you know, ceilings, you know, [music] insulations.

[00:15:18]>> I see.

[00:15:18]>> So it's again the design and it's still [music] in the process of getting the design done.

[00:15:24]>> Uh that that's it. That was it for me.

[00:15:26]Do you Did you want to share anything else that maybe [music] we have missed?

[00:15:30]>> Uh One thing I want to share is uh uh don't be too optimistic about the you know the autonomy of the of the human uh annoyed robot.

[00:15:40]Uh because uh challenges fundamental challenges are still there. [music] You you can't just put uh large language models uh on it yet.

[00:15:48]>> Mhm.

[00:15:48]>> Because there are interesting constraints. [music] And there are like basic physics constraint from the thermal side.

[00:15:54]>> Right.

[00:15:55]>> Uh people haven't realized it, but soon it will be a a challenge. [music] So but uh again the don't be too pessimistic as well. Because uh >> [music] >> even though uh you can achieve a full autonomy, it can be useful already.

[00:16:11]Because uh with like I mentioned, the teleop the remote teleop uh you know maturity, it it will come soon that >> [music] >> uh you can just remote control the robot in a nuclear plant.

[00:16:23]>> Mhm.

[00:16:23]>> Which already has uh some >> [music] >> certain values. Right. You can use the robot to uh you know uh work in the you know human not friendly environments. For example, in the uh rubber company uh rubber factory.

[00:16:37]Because when there is it's smelly, it's it's hard, it's you know. So you you already have uh a chance to use humanoid in the human environment. So uh you know just keep a uh know realistic view and you'll find the values of the humanoid.

[00:16:53]>> Just curious what's like the good use case you're expecting for Ollie and Shaun to use?

[00:17:00]>> Okay. Yeah, I it's it's it's a separate.

[00:17:03]For Ollie, like I said, it's it's not uh reliable uh deployable uh humanoid yet. But it can be used uh uh to uh you know in for example for example entertainment, >> [music] >> uh tourism, you know, concierge, this kind of environment [music] which don't need a lot of precise manipulation by the you know certain kind of interaction of the human being which [music] is doable at this moment. So, I think that the big use case is there and from there people get used to the humanoid working with the human and then later on with a better, you know, optimization optimized design, it can be gradually moved to the [music] you know household and the factories.

[00:17:51]So, for Tron it's a different story.

[00:17:53]It's basically a all terrain platform.

[00:17:57]It's basically an extension from the previous AGV AMR >> [music] >> type of platform which can only work on, you know, flat grounds. Now, you have a upgrade.

[00:18:09]It can work on the different [music] difficult terrains and then you can put your own payload on top of it. So, this is a different philosophy and we see that we're we're already solved the locomotion problem on Tron so that we expect our [music] users can, you know, use their imaginations and put different payloads on on Tron to have different [music] applications from there. Yes.

[00:18:34]>> Will there be a new generation of Olli or Tron?

[00:18:38]>> Yeah, yeah, in a couple of life.

[00:18:41]>> It's very like >> different >> different line with a different >> We are actually >> [laughter] >> No, we're building on our experience.

[00:18:51][music] So, you will see a variation from it, but it's an improvement. It's not a radical >> [music] >> change of a design.

[00:18:58]>> I'm looking forward for more Olli >> No problem.

[00:19:00]>> and Tron in the world.

[00:19:01]>> You'll see in a few months. [music] Yes.

[00:19:04]Thank you. Thank you.

[00:19:09]>> [music] >> Woo!